A Conceptual Framework for Shared Autonomy

Real human-AI systems do not operate at a single level of autonomy. Their effectiveness depends critically on how, when, and under what engagement conditions the human operator intervenes. Inspired by the Pilot Authorization and Control of Tasks (PACT)style paradigms, we present a decision-theoretic framework for mixed-initiative interaction under shared autonomy. The key idea is to treat the interaction context as an object relevant for planning, rather than a collection of ad-hoc autonomy modes and hard-coded constraints. At a high level, our framework enables reasoning about richinteractionprotocolswithinaMarkovianplanningmodel, withoutenumeratingautonomymodesorrelyingonrecursivemodelsof human reasoning. We prove that the resulting formulation reduces to a standard Markov decision process (MDP), while remaining compact and scalable, and admits solution via classical MDP solvers such as value iteration. Empirically, this structure enables qualitatively richer policies that adapt not only to the task state but also to evolving interaction context, which yields behaviors beyond the scope of the models available today. All formal proofs and experimental details are deferred to a longer version.

Authors

• Shashank Shekhar Université de Caen Basse Normandie
• Laurent Jeanpierre Université de Caen Basse Normandie
• Abdel-illah Mouaddib Université de Caen Basse Normandie

Keywords

• Human-AI teaming
• Mixed-initiative planning
• Joint actions
• MDPs